You probably did not look at every word in this sentence for exactly the same amount of time.
Your eyes may have moved quickly over the predictable parts, paused slightly longer somewhere else and perhaps even jumped backwards when something stopped making sense.
That is not necessarily poor reading.
New research suggests these movements may form part of a remarkably efficient system in which the brain continuously decides where looking again is worth the time and effort.
Researchers have developed a computational model that treats reading as a sequence of decisions made under limited attention, memory and time.
The model reproduced a wide range of familiar human reading behaviours, including why we spend longer looking at unusual words, why predictable words are sometimes skipped and why our eyes return to earlier parts of a sentence when comprehension begins to break down.
Published in Nature Human Behaviour, the study proposes that reading can be understood as a form of resource-rational decision-making.
In simpler terms, your brain may be constantly asking one question while you read:
Is looking at this again worth the effort?
Your eyes move far more than you realise
Reading feels smooth because we experience the meaning of a sentence rather than the mechanics involved in extracting it.
Our eyes actually move through text in a series of short jumps and pauses.
The pauses are known as fixations.
During a fixation, visual information is collected from the word or region around which the eyes have stopped.
The rapid movements between these points are called saccades.
Readers also regularly skip words entirely and sometimes make regressions, moving their eyes backwards to something they have already read.
These behaviours are not random.
Decades of eye-tracking research have shown that people tend to spend longer looking at words that are longer, less common or harder to predict from the sentence around them.
Short, frequent and highly predictable words are more likely to receive less attention or be skipped.
The difficult question has been explaining how all these individual eye movements connect to the larger goal of actually understanding the text.
The researchers treated reading like a decision problem
Yunpeng Bai and colleagues approached the problem differently.
Instead of programming a model with rules such as “skip short words” or “reread difficult sentences”, they gave it a broader objective.
The simulated reader had to maximise comprehension while paying costs for time, memory use and the effort involved in gathering additional visual information.
The model then had to learn how to read efficiently under those constraints.
This is known as resource rationality.
The idea is that human behaviour does not need to be perfectly optimal in a world of unlimited time and memory.
It needs to make sensible use of the limited resources actually available.
When applied to reading, that means looking at every single letter for as long as possible would not necessarily be the best strategy.
It might maximise information, but it would make reading painfully slow.
The better strategy is to collect enough information to understand the text without wasting attention where additional effort is unlikely to help.
The model learned when a word was worth looking at
Consider the sentence:
She poured herself a cup of…
By the time your eyes reach the final word, the context has already narrowed the possibilities considerably.
If the next word is “coffee”, a reader may need less visual information to identify it than if the sentence suddenly ends with an unexpected word.
The model behaved in a similar way.
When a word was frequent or highly predictable from its context, uncertainty about its identity was already relatively low.
The additional benefit of staring at it for longer was therefore small.
Harder or less predictable words produced more uncertainty, making additional visual attention more valuable.
From this single principle, the model reproduced the well-established finding that readers spend more time on long, rare and unexpected words.
Skipping a word is not necessarily laziness
The same logic helps explain why readers sometimes skip words.
If enough information is already available from context and peripheral vision, directly fixing your eyes on a predictable word may add very little.
Moving on can therefore be efficient.
The researchers found that their model reproduced human patterns of word skipping without being explicitly programmed to imitate them.
Shorter, more common and more predictable words were skipped more frequently.
But skipping became less attractive as sentences became more ambiguous.
When several interpretations were possible, the value of gathering additional information increased.
This provides a useful distinction between efficient reading and careless skimming.
Skipping information can make sense when the reader already has enough evidence to maintain comprehension.
Skipping becomes costly when the missing information is necessary to understand what follows.
Going backwards can be a sign that your brain noticed a problem
Readers also make regressions.
Your eyes reach a later word, something does not fit, and they jump backwards.
It can feel like a mistake.
The new model suggests it may instead represent an intelligent response to uncertainty.
When the current interpretation of a sentence became difficult to integrate, returning to an earlier word provided another opportunity to gather useful information.
The model consequently made more regressions when words or sentences were difficult or ambiguous.
At the level of an entire text, it also returned to previous sentences when comprehension became less coherent.
Rereading, from this perspective, is not simply repeating work.
It is targeted information gathering.
Your reason for reading changes the way you should read
One of the more interesting implications is that there may not be a single ideal reading speed.
The best strategy depends partly on what you are trying to achieve.
If the goal is deep comprehension, missing an important relationship between two paragraphs carries a large cost.
Spending additional time rereading may therefore be worthwhile.
If you are searching a document for one name or quickly checking whether an article contains relevant information, the value of understanding every sentence is much lower.
Reading faster and tolerating incomplete comprehension can become perfectly rational.
The researchers’ simulations changed in exactly this way.
When comprehension was given greater value, the model became more careful and information-seeking.
When comprehension was less important relative to time and effort, it adopted faster strategies and accepted a less complete understanding.
Time pressure changed how real people read
The researchers did not rely only on computer simulations.
They also collected a new dataset in which participants read English texts under different levels of time pressure while their eye movements were recorded.
As the available reading time decreased, participants adapted.
They read faster, skipped more material and made fewer backward eye movements.
There was a cost.
Performance on multiple-choice comprehension questions and free-recall measures declined as time pressure increased.
Human eye-movement measures in this part of the study were based on 28 participants, while comprehension results were calculated from 32 participants.
The computational model reproduced the same broad shift between speed and understanding.
That is important because it suggests that faster reading is not simply the same process running at higher speed.
Readers may actively change how they allocate attention when time becomes scarce.
This explains why speed-reading has an obvious limit
The findings provide a useful way to think about claims that people can dramatically increase reading speed without sacrificing comprehension.
There are genuine efficiencies available during reading.
You do not need to process every word equally.
Context allows some information to be predicted, and rereading can be focused on the places where uncertainty is greatest.
But there is still a trade-off.
Reading requires information.
If speed increases mainly by reducing the amount of information sampled from the page, comprehension will eventually suffer.
The human time-pressure experiment showed exactly this pattern: more skipping and fewer regressions were accompanied by poorer recall and comprehension.
The brain can become more economical.
It cannot obtain unlimited information from text it never processes.
This matters when studying textbooks
The research also offers a useful lesson for students.
A difficult textbook should not necessarily be read at the same pace as a familiar article.
Prior knowledge changes how predictable information is.
Someone studying a subject for the first time has more uncertainty to resolve and may therefore need more fixations and more rereading.
A student who already understands the underlying concepts can process the same paragraph with less effort because much of its structure is already predictable.
This helps explain a familiar experience in higher education.
A page of material in a subject you know well can disappear in a minute.
A page containing unfamiliar mathematics, chemistry or theory can take ten.
The second reader is not necessarily worse at reading.
The text simply demands more information before a coherent understanding can be built.
Rereading the difficult sentence may be exactly what you should do
Students are sometimes encouraged to keep moving when reading becomes difficult.
There is value in maintaining momentum, particularly when excessive rereading becomes unproductive.
But the new model provides a useful defence of strategic regression.
If something later in the paragraph reveals that your current interpretation is wrong, going backwards can be the most efficient next action.
The important distinction is whether rereading is targeted.
Returning repeatedly to an entire page without knowing what went wrong may achieve little.
Returning specifically to the sentence that introduced the confusing concept can supply the information needed to repair understanding.
The difference is subtle but important.
Effective reading may involve knowing not only when to continue, but when continuing is actually the less efficient choice.
This is still a model of reading, not a scan of the brain
There are important limitations.
The researchers developed a computational explanation of reading behaviour. The fact that the model reproduces many human patterns does not prove that the human brain literally performs the same calculations internally.
The model was designed primarily to explain behaviour at the group level rather than every individual reader.
Real people differ in vocabulary, language background, reading ability, prior knowledge and personal strategies.
The new experimental dataset was also relatively small and focused specifically on English reading under time pressure.
Further studies will need to examine a wider range of languages, genres, text lengths, reading goals and populations.
The researchers also measured comprehension mainly through recall and question accuracy.
Real reading can have very different goals, including enjoyment, emotional engagement, long-term learning, information search and decision-making.
Good reading may be less about seeing everything than knowing where to look
We often imagine a strong reader as someone whose eyes move smoothly from the first word of a page to the last.
Real reading is considerably messier.
Eyes stop, jump, skip and sometimes turn around.
The new research suggests that messiness may be part of the intelligence of the system.
Attention is limited.
Memory is limited.
Time is limited.
A capable reader therefore cannot simply collect every possible piece of information.
They have to decide which information is worth collecting.
Sometimes that means lingering on one difficult word.
Sometimes it means skipping something obvious.
And sometimes the smartest way to move forward through a book is to move your eyes backwards first.
Source Information
Study Title: Hierarchical resource rationality explains human reading behaviour
Authors: Yunpeng Bai, Xiaofu Jin, Shengdong Zhao, Antti Oulasvirta et al.
Journal: Nature Human Behaviour
Published: 10 August 2026
Human time-pressure study: 28 participants for eye-movement measures and 32 participants for comprehension measures
DOI: 10.1038/s41562-026-02534-0
